Deep Learning-Enhanced Industrial Visual Inspection System with Multi-Agent Collaboration for Manufacturing Quality Assurance

Abstract

Quality assurance is one of the most critical stages in modern manufacturing, directly affecting product reliability and safety. Traditional visual inspection relies heavily on human inspectors, which is labor-intensive, subjective, and difficult to scale in high-throughput production lines. Recent advances in deep learning and multi-agent systems offer new possibilities for automating and enhancing visual inspection accuracy. This study proposes a Multi-Agent Deep Learning Visual Inspection System (MADL-VIS) for industrial manufacturing quality control. The system employs a two-stage detection pipeline: a deep learning-based defect detection module that identifies surface defects from visual inputs, and a multi-agent collaboration module that performs defect classification, root cause analysis, and inspection report generation. Specifically, the detection module leverages a convolutional neural network architecture with attention mechanisms to extract fine-grained defect features from product surface images. The multi-agent module decomposes the post-detection workflow into specialized tasks—defect categorization, severity scoring, cause inference, and documentation—each handled by a dedicated LLM-powered agent. Experiments conducted on three publicly available industrial inspection datasets demonstrate that the proposed system achieves an average defect detection accuracy of 91.3% and a classification accuracy of 88.7%. The multi-agent post-processing module reduces the average inspection cycle time by 38% compared with manual analysis, while the root cause inference agent achieves a consistency rate of 79% with domain expert assessments. This study validates the effectiveness of integrating deep learning-based visual detection with multi-agent collaborative analysis for automated industrial quality assurance.